Because this isn't training an audio model along with a million other voices to understand English, etc. It's clearly meant to sound exactly like that one celebrity.

I suspect a video avatar service that looked exactly like her would fall afoul of fair use as well. Though an image gen that used some images of her (and many others) to train and spit out generic "attractive blonde woman" is fair use in my opinion.

Chances are this is. Basically same as LoRA. One of go-to tools for these literally uses Diffusion model and work on spectrograms as images.

Okay so as long as we steal enough stuff then it's legal.